Which text features best predict mental health from language?

The paper argues that text-based prediction of mental health outcomes "rarely specif[ies] what they measure," and proposes NLP Psychometrics: treat psychological prediction from text as a psychometric problem, tie each score to interpretable linguistic evidence, and test beyond the training text format. Its first headline finding is that "language carried most of the recoverable psychometric signal." Emotion and network features formed a predictive core across life satisfaction (SWLS), depression (PHQ-9) and DASS-21, while sociodemographics alone explained no meaningful variance in depression, anxiety or stress. Life satisfaction was the exception, where emotion features and income were the strongest predictors.

The paper's mechanism is a measurement premise. Inner experience leaves quantifiable traces, some explicit (questionnaire item scores) and some linguistic, "distributed across the words people choose, the emotions they express, and the concepts they connect." Building on the mental lexicon as a network of interconnected concepts, it extracts emotional profiles and syntactic-semantic structure as textual forma mentis networks. Those features go into ablated random forest regressors alongside personality and sociodemographic variables, and SHAP shows which features drove performance and in which direction. The full models explained up to 70.8% of variance in SWLS, 55.7% in PHQ-9, and 68.5%, 76.0% and 72.4% for DASS-21 depression, anxiety and stress. The markers were "interpretable and construct-specific": neuroticism and network topology dominated depression and anxiety, reversing direction between them.

This sits close to Why do discourse patterns predict anxiety better than single words?. Both say that for depression and anxiety the informative signal lies in how concepts connect, not in isolated vocabulary. The paper adds a different operationalization (network topology, not annotated discourse relations) and a competing predictor class, personality. It also adds a contrast between constructs, since life satisfaction leans on affect and income while depression and anxiety lean on neuroticism and topology. It reads alongside Can language summaries unlock hidden psychological patterns?, where trait scores and language summaries are related in the opposite direction. It also reads alongside Can local language models rate therapy engagement reliably?, another case of LLM-produced text serving as psychometric material.

The excerpt leaves a good deal open. The text came from nine LLMs "conditioned on controlled personas" (cognitive digital shadows) that completed questionnaires with per-item explanations, so the headline variance figures describe machine-generated language, not patient language. The third finding says the mapping transferred to LLM-generated diaries and to speech transcripts of clinically depressed patients and controls, but only "with reduced yet significant accuracy." The excerpt gives no transfer figures, sample sizes or persona construction details. The safe reading is that linguistic features are a promising, interpretable measurement layer for scores, and that clinical validity on human data remains a separate, weaker result to check in the full paper.